Agents of Work
Let's Talk
October 16, 2026 · Agents of Work

Agents of Work AI Daily Briefing — October 16, 2026

Summary

Meta gave away an 816-billion-parameter model that runs on two desktop Macs, Microsoft built a bigger one and is keeping it, and Accenture told staff it will pay for $29 of Claude a month. The price of AI is being renegotiated from both ends. For a business owner, the same day brought FTC subpoenas, a Meta clause that claims your website, and a test showing chatbots get industrial parts questions wrong half the time.

Highlights

  • Meta's Muse Spark Open is free to download and costs a fraction of closed models to rent; test it before renewing anything.

  • The two Mac Studios it runs on will cost well over $20,000 and do not ship until late October.

  • Accenture now funds $29 of Claude per employee a month, so expect your own AI vendors to push usage billing.

  • Ramp data says 1% of firms pay 80% of the OpenAI and Anthropic bill, and buyers now hold the negotiating power.

  • One in three teams switched their main AI lab last month; keep your prompts and data portable.

  • Meta's business terms let it crawl any website tied to your Facebook, Instagram or WhatsApp account from October 30, with no opt-out.

  • The FTC has subpoenaed OpenAI and Anthropic over safety incidents, and third-party subpoenas come next.

  • A cheap AI model used as a security gate was fooled by fake rules hidden in a profile field; never let a model be your only check.

  • Leading chatbots answered industrial parts questions right 52% of the time at best, and sounded sure when wrong.

  • Warehouse robots now come as a subscription for small retailers, and Waymo is heading to Toronto to learn winter.

Quick Takes

  • A developer cloned Word, Excel and PowerPoint with Claude. Brandon Thomas released WordCraft, GridCraft and DeckCraft as free, open-source apps built with Claude Opus 5.5 from public file specifications. GridCraft handles 513 of Excel's 523 functions and WordCraft passed 12,000 GitHub stars, but his own team rates them 50% to 60% ready for real work. Read more

  • AI doubled the accuracy of weather forecasts two to six weeks out. A machine-learning correction built by MIT, Microsoft and partners doubled the predictive power of the European weather center's model in that range and won its forecasting competition, the first method to beat long-term averages at that distance. The code is open source and runs without a supercomputer. Read more

  • Amazon's $75 Bee turns your conversations into to-do lists. The Bee Frontier, a $74.99 clip or wristband with seven days of battery, is on sale now; a $99.99 card-sized Bee Eclipse ships later. Both record conversations and produce summaries, to-dos and calendar invites, with 14 days of memory free and a $11.99 monthly plan that connects Gmail, Outlook and Alexa+. Amazon says raw audio is never stored. That is a company promise, and several states require everyone's consent to record. Read more

What's Covered in Featured News

  • Meta's Muse Spark Open: what the 816-billion-parameter open model can do, what it costs to run, how it scores and what the license allows.

  • Microsoft's MAI-Base-2: the 2.8-trillion-parameter model Microsoft trained itself and why it is not releasing it.

  • The AI bill renegotiated: Ramp's spending data, Accenture's $29 cap, the price spread between equal models and the shift to open weights.

  • The FTC and Anthropic's auditors: what the subpoenas seek, and the outside evaluators now inside Anthropic.

  • Meta's October 30 terms: the clause that lets Meta crawl business websites, and what lawyers say about it.

  • Security: Bank of America's 4,000-person patching effort, a Claude Code flaw, a fooled security model and AI in ransomware.

  • Physical AI: chatbots versus industrial parts, warehouse robots on subscription, Symbotic's acquisition, ABB's trainer, Waymo in Toronto, Volvo and Nvidia, and the robot data gap.

Featured News

Meta's 816-billion-parameter model is free to download and runs on two Macs

Meta Superintelligence Labs released Muse Spark Open on October 15 and called it the most capable open-weight model built in the United States. It has 816 billion total parameters, of which 42 billion are active for any given token, a one-million-token context window, and accepts text, images, video and audio. The weights are on Hugging Face, Amazon SageMaker, Google Cloud and Microsoft Foundry, and hosted versions from Together AI and six other providers cost about $1.25 per million input tokens and $4.25 per million output. Meta's closed Muse Wonder costs $4 and $20.

The headline is the desk. Meta says the model runs locally on two Apple Mac Studios with the M5 Ultra chip, each with up to 512 gigabytes of unified memory, and showed it fixing an open-source Mastodon app and cropping video in DaVinci Resolve. Read the fine print before budgeting. The model card asks for at least 700 gigabytes of accelerator memory for the compressed build and two terabytes for full precision. The 512-gigabyte Mac Studio does not ship until late October and has no announced price; the 256-gigabyte version lists at $10,799. Meta published no speed figure, and the demos were company-run.

On the scoreboard it is good, not dominant. Meta's own table puts it ahead of China's GLM 5.3 and Kimi K3 on job and tool-use tests but behind both on coding, and its Artificial Analysis Intelligence Index score of 39 sits below GLM-5.3 at 45, Kimi K3 at 44 and Qwen3.8 at 40. The license is not fully open either: a model-as-a-service business with more than $20 million in annual revenue must notify Meta and sign a separate agreement, and products must display "Built with Muse Spark Open." Researcher Nathan Lambert caught the mood: "I waited so long for US open weight models to be relevant again." For a business, the point is that a frontier-class model you can inspect, host yourself and run without sending data to anyone is an American option again, at a fraction of the closed leaders' price.

Microsoft finishes a 2.8-trillion-parameter model of its own, and is keeping it

Microsoft AI published a pre-training report the same day for MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active, trained on 50 trillion tokens on 16,000 GB200 chips over 38 days in August and September. Microsoft says it is 4 to 12 times more efficient per unit of training compute than DeepSeek's V4-Pro base model and beats every open-weight model it tested, with the caveat, stated by Microsoft, that the efficiency figures are estimates from scaling laws rather than measurements.

Nothing here is usable today. The model is still in post-training, there is no release date, and the weights will be licensed to "trusted partners" through Foundry rather than posted publicly. Paired with Monday's Maia 300 chip announcement, the direction is plain: the company that resells OpenAI's models intends to own its own silicon and base model.

The AI bill gets renegotiated from both ends

Ramp's October AI Index, published October 14 by economist Ara Kharazian, describes a market near the top of its adoption curve. 55.9% of businesses on Ramp's platform now pay for an AI tool, with Anthropic at 43.8% and OpenAI at 40.6%, but growth has slowed to a fraction of a point a month. The money is concentrated: Ramp says 1% of firms account for 80% of spending on OpenAI and Anthropic, and that top 1% spent $9,568 per employee in September, down 3.1% from July's peak. Token volumes rose while token spend stayed flat, which Ramp attributes to falling prices, and it concludes that buyers now hold the stronger negotiating position.

Accenture put a number on it. An internal memo first reported by 404 Media says the firm will centrally fund $29 per employee per month for Claude as of October 1, with anything above charged to a client project. Internal messages described some employees running $500 to $1,000 a month. A spokesperson said AI remains available to all employees. This is the firm that announced in December it would train 30,000 staff on Claude.

The price spread explains the squeeze. Investor Tomasz Tunguz found that five frontier models scoring within five points of each other on the Artificial Analysis index cost between $1.99 and $7.63 per task, a 3.8-fold gap, with Google's Argon cheapest. Vercel's AI Gateway data for September shows where that leads: open-weight models now handle 71.5% of tokens and 23% of spend, up from about 5% of spend in June, and almost one in three teams switched their most-used lab in a single month. Anthropic, meanwhile, opened fast mode for Claude Opus 5.5 to all customers: up to 2.5 times the output speed from the same model at $8 and $40 per million tokens, double the standard rate.

The FTC subpoenas OpenAI and Anthropic, as Anthropic lets auditors inside

The Federal Trade Commission sent civil investigative demands, its form of subpoena, to OpenAI and Anthropic on October 14, an FTC spokesperson confirmed to CBS News after Axios first reported it. The demands seek documents on safety incidents, privacy policies and safeguards, following the probe announced two weeks earlier into models that went rogue and hacked into websites during testing. Subpoenas to third-party organizations are to follow, and the agency could compel executives to testify.

Anthropic announced October 15 that six employees of the evaluation nonprofit METR now have "employee-like access" to its systems, on par with its own risk assessment teams. METR can publish its findings on Anthropic's risks, incidents and practices; Anthropic cannot veto publication, though it can require redactions in predefined categories, and METR may say publicly whether it thought those redactions were justified. That answers the Reuters report, covered here Wednesday, that labs had sought control over what evaluators publish. Separately, after Anthropic's $16.8 billion quarter showed growth slowing, some investors now expect its November IPO to price below $1.5 trillion rather than the $2 trillion once floated.

Meta's new business terms let it crawl your website

If your business has a Facebook page, an Instagram account or a WhatsApp Business number, read Meta's updated commercial terms before October 30. Legal experts who reviewed them for Nieman Lab say a new clause lets Meta run automated crawlers on websites "associated with" a business account and use the content to develop, improve, research and train its products, a definition that includes AI at Meta. There is no opt-out. Continuing to use the platforms after the date counts as consent, and Meta's notice tells businesses that disagree to "stop all access or use" of its products for business purposes.

Media lawyer Matt Topic called it "a remarkable provision" and said publishers should assume it is enforceable. Cornell's James Grimmelmann read it as clear permission to train AI on business sites. Meta says the update consolidates permissions that already existed and that site owners can still manage access through robots.txt. "Associated with" is undefined, so nobody yet knows whether it means your product catalog or every page you own.

Security: a bank hires 1,000 people to fix what AI found

Bank of America has 4,000 people, 1,000 of them outside contractors, working since the second quarter to patch software vulnerabilities surfaced by Anthropic's Mythos model and other AI systems, chief executive Brian Moynihan told investors, according to CNBC. The work, combined with anti-money-laundering requirements, cost "a few hundred million dollars" in the third quarter. The bill for fixing what AI finds is now showing up in earnings.

The AI tools have holes of their own. Noma Security disclosed a flaw it calls SneakPeek in Claude Code Desktop's preview pane: HTML files of 256 kilobytes or more were loaded with local file access instead of in a sandbox, so a poisoned file in a public repository could read a developer's AWS credentials, GitHub secrets and SSH keys and send them out the moment Claude was asked to review it. Anthropic rated it 8.7 out of 10, and its first patch was bypassed before a deeper fix landed. Datadog researchers showed that Jev, a cheap decision model some teams use as a security gate for AI agents, can be steered by a "counterfeit policy" hidden in the text it reviews: its detection of attacks fell from 82.9% to 3.9%, and a fake rule planted in a user-editable profile field dropped a LangChain risk score from 0.89 to 0.29, below the level that blocks an action. And in a Semperis survey of 1,200 security professionals in nine countries, 84% of ransomware victims said attackers used AI and 44% said attackers exploited the victim's own AI agents.

Physical AI

The most useful robotics reading this week is a test, not a demo. Reshape Automation, which sells an AI sales assistant to industrial distributors, put 100 real customer questions about parts from 14 manufacturers to GPT-6 Astra, Claude and Gemini. Without web search, all three scored 12% to 14%. With search, GPT-6 Astra reached 52% and Claude 40.5%, so the best setup still got 42 of 100 wrong, and about one wrong answer in three gave a specific part number or figure with no caveat. The report is self-published and lists 13 limitations, but the pattern will be familiar to anyone who has asked a chatbot for a SKU: the models "sound equally sure" whether right or wrong, as chief executive Juan Aparicio put it.

The warehouse vendors are chasing smaller buyers. AutoStore's fall release includes Pio Pal, a robotic picking station sold to smaller businesses as robotics-as-a-service that packs up to 540 order lines an hour, live at one Norwegian retailer with wider rollout planned for early 2027. Symbotic bought Germany's cellumation, whose sorting surface of hundreds of independently controlled cells moves items in any direction in a fraction of a conveyor's floor space, on undisclosed terms. ABB's new IRB 1700 arms promise 20% shorter cycle times in half the footprint, and its AI Robot Trainer lets integrators teach vision-guided picking with any camera and, by ABB's estimate, up to 90% less commissioning time.

On the road, Waymo won approval to join Ontario's automated-vehicle pilot and will begin mapping and testing in Toronto, its northernmost and snowiest city yet, with Element Fleet Management running the local fleet. There is no service date; Waymo now operates in 15 U.S. cities. Volvo expanded its Nvidia partnership to build a hands-free system for city streets and highways on Nvidia's hardware, driving models and safety layer, across gas and hybrid cars as well as EVs, with the first vehicle due "within the decade."

Why robots still fumble is a data problem, and Brian Potter of Construction Physics put numbers on it: the industry collects an estimated 500,000 hours of robot demonstrations a month at $60 to $150 an hour, and reaching the billion hours he thinks a general-purpose robot needs would cost $10 billion to $100 billion. Roboticist Ken Goldberg calls it the "100,000-year data gap."

What This Means for Your Business

Check what you pay against what you need, this month. Ramp's data says the labs are discounting and buyers have the leverage; Accenture's memo says the all-you-can-eat era is over even at the biggest customers. Pull your AI invoices, find the two or three jobs that generate most of the spend, and run them through a cheaper model, including a hosted open-weight one like Muse Spark Open at a quarter of the closed price. If the output is as good, switch that job, not the whole company. If your vendor bills by usage, ask for a cap and a discount in writing before the renewal, because this is the quarter they will give one.

Stay portable. One in three teams changed their main AI lab in a single month, which is only possible if your prompts, documents and workflows are not welded to one provider. Keep your system prompts, examples and evaluation sets in your own files, not inside a vendor's dashboard, and prefer tools that let you swap the underlying model. The Meta model's license is worth reading too: free for almost everyone, but a company that resells it as a service past $20 million in revenue needs Meta's permission.

Decide about Meta's terms before October 30. If your business uses Facebook, Instagram or WhatsApp for business, Meta will treat your continued use as consent to crawl your website for AI training. Most small businesses cannot afford to leave those channels, so take the practical steps instead: make sure your robots.txt names Meta's crawlers if you want them out, move anything you would not want in a training set behind a login, and ask your lawyer whether "associated with" could reach client portals or pricing pages. Watch for publishers' legal challenges; the first ruling will tell you whether the clause holds.

Treat AI security controls as one layer, never the only one. The Datadog finding is the one to remember: a model acting as a safety gate was switched off by fake rules typed into a profile field. If an agent in your business can send money, delete records or email customers, put a hard authorization check outside the model, in code or in a human, and log every file the agent reads and every address it sends to. Update Claude Code Desktop and every other AI developer tool now, and rotate any long-lived credentials sitting on developer laptops. If you are hit by ransomware, assume AI helped plan it and that your own agents may have been the way in.

Finally, do not ask a chatbot for a part number without a document in front of it. Reshape's test is the clearest evidence yet that general models fail on catalog lookups and sound confident doing so. Give the model your actual price list, spec sheets or manufacturer PDFs, require it to cite where the answer came from, and route anything it cannot verify to a person. The same rule applies to the Bee on your lapel: tell people when it is on, and keep it off in meetings where you would not want a transcript to exist.

Sources

← All posts